Comparative analysis of one- and two-dimensional spectral representations for LIBS-based classification of rocks

Laser-induced breakdown spectroscopy (LIBS) demonstrates substantial potential for rapid rock identification. However, its classification performance is limited when distinguishing rocks with highly similar elemental compositions due to the high dimensionality of spectral data and complex spectral feature relationships. This study employed seven diorite-related rock samples to systematically evaluate the effects of one-dimensional spectral representation and two-dimensional spectral transformation on classification performance. First, the original spectra were resampled from 16 376 to 1024 wavelength points using the piecewise cubic Hermite interpolation polynomial (PCHIP). Support vector machine (SVM) and one-dimensional convolutional neural network (1D-CNN) served as baseline models for one-dimensional classification. Subsequently, the one-dimensional spectra were converted into two-dimensional spectral images using Gramian Angular Summation Field (GASF) and continuous wavelet transform (CWT). Four two-dimensional models—2D-CNN, DenseNet121, GoogLeNet, and MSCNN—were then applied for classification. The results demonstrated that two-dimensional spectral representation significantly improved classification performance compared with direct one-dimensional spectral classification. The classification accuracies of SVM and 1D-CNN were 74.76% and 82.86%, respectively, whereas the accuracy, precision, and recall values of all two-dimensional models under both transformation methods exceeded 85%. Furthermore, the results indicated that the compatibility between model architecture and spectral transformation method further influenced the final classification performance. DenseNet121 achieved the highest accuracy of 97.62% under the GASF representation. In contrast, GoogLeNet and MSCNN performed better under the CWT representation, attaining accuracies of 95.71% and 96.67%, respectively. These findings suggest that two-dimensional spectral representation enhances the capability of LIBS to discriminate between similar rock samples. Furthermore, the final classification performance depends not only on the transformation method but also on its compatibility with the model architecture. This study provides a reference for optimizing spectral representation strategies in LIBS-based rock classification.